JP6148150B2 - Acoustic analysis frame reliability calculation device, acoustic model adaptation device, speech recognition device, their program, and acoustic analysis frame reliability calculation method - Google Patents

Description

  The present invention relates to an acoustic model adaptation technique used for speech recognition, and in particular, an acoustic analysis frame reliability calculation apparatus and method for selecting feature quantities used for unsupervised acoustic model adaptation, and an acoustic model using the apparatus. The present invention relates to an adaptive device, a speech recognition device, and their programs.

  When updating the acoustic model used for speech recognition, model parameter optimization processing is performed so that as many examples as possible are included in the learning data. This process is referred to as “acoustic model adaptation” and generally uses an audio file, correct text representing the utterance content of the audio file, and learning (adaptive) data. The adaptation of the acoustic model is roughly divided into two types: supervised adaptation in which correct text is obtained by human writing up the reading corresponding to the speech file, and unsupervised adaptation obtained as a speech recognition result of the speech file.

  Unsupervised adaptation is superior in terms of cost and time because it does not involve human intervention, but the speech recognition results may include misrecognition, so the adaptation process reduces the accuracy of the acoustic model There is. For this problem, a confidence measure indicating the reliability is given to the speech recognition result, adaptive data is selected according to the height of the confidence measure, and the acoustic model is adapted using the selected speech recognition result. A method is considered (Patent Document 1).

  This method assigns a confidence measure to the speech recognition result of each utterance in the speech file, and selects only utterances whose confidence measure exceeds a certain threshold as adaptation data to perform unsupervised adaptation. Here, the utterance is a voice section of several seconds to several tens of seconds uttered by, for example, one breath in the voice file, and the speech recognition result of the utterance usually includes several words to several tens of words.

JP 2007-248730 A

  The prior art is a method in which an utterance having a low confidence measure, that is, having a low speech recognition rate includes many misrecognitions, and does not select the speech recognition result of the utterance as adaptive data. Even for an utterance with a low confidence measure that is not selected as adaptive data, the entire utterance section is rarely misrecognized, and often includes a correct recognition result.

  However, in the prior art, since the utterance is selected as a unit, it is not possible to exclude only the erroneous recognition section. In short, in the conventional technology, the erroneous recognition section cannot be accurately identified. As a result, there has been a problem that the cost of collecting the adaptive data is increased, the adaptation effect is lowered, and the improvement of the speech recognition accuracy is reduced.

  The present invention has been made in view of such a problem, and an acoustic analysis frame reliability calculation apparatus and method for accurately identifying an erroneous recognition section in adaptive data, and an acoustic using the apparatus. The object is to provide a model adaptation device, a speech recognition device, and their programs.

  The acoustic analysis frame reliability calculation device according to the present invention includes a speech recognition unit and a frame reliability scale calculation unit. The speech recognition unit divides the speech signal into acoustic analysis frames having a predetermined time length, extracts acoustic features in units of the acoustic analysis frames, and calculates acoustic likelihood using the acoustic features and the initial acoustic model. Then, the maximum likelihood phoneme ID is assigned to the acoustic analysis frame, and a phoneme ID / acoustic likelihood acoustic feature quantity sequence in which the acoustic feature quantity, the acoustic likelihood, and the phoneme ID are associated with each other is output. The frame confidence measure calculation unit receives the phoneme ID / acoustic likelihood-equipped acoustic feature amount sequence output from the speech recognition unit, classifies the acoustic feature amount for each phoneme ID, and creates an acoustic feature amount set, Outlier detection is performed on each acoustic feature quantity set to obtain an outlier score, a function value using the outlier score is obtained as a measure of reliability, and the value of the above function is assigned to each acoustic analysis frame The acoustic feature quantity series with frame reliability is output.

  The acoustic model adaptation apparatus of the present invention includes the above-described acoustic analysis frame reliability calculation device, a feature amount selection unit, and an acoustic model adaptation unit. The feature amount selection unit receives an acoustic feature amount sequence with frame reliability output from the acoustic analysis frame reliability calculation device as an input, and selects a sound analysis frame having a reliability for each acoustic analysis frame equal to or greater than a selection threshold. The acoustic feature quantity series with selection flag to which the selection flag shown is given is output. The acoustic model adaptation unit uses the acoustic feature quantity sequence with the selection flag output from the feature quantity selection unit as an input, and the statistical amount calculated using the acoustic feature quantity of each acoustic analysis frame to which the selection flag is not assigned is 0 or more. An updated statistic is obtained by multiplying a non-selection weight of a real value of 1 or less, and the parameters of the initial acoustic model are updated based on the updated statistic, and an after-adaptation acoustic model is output.

  The speech recognition apparatus of the present invention includes the above-described acoustic analysis frame reliability calculation device, a feature amount selection unit, and a speech recognition unit. The feature quantity selection unit indicates that an acoustic analysis frame with frame reliability output from the acoustic analysis frame reliability calculation device is input, and an acoustic analysis frame having a reliability for each acoustic analysis frame that is equal to or higher than a selection threshold is selected. The acoustic feature quantity series with the selection flag to which the selection flag is assigned is output. The speech recognition unit inputs the acoustic feature quantity sequence with the selection flag output from the feature quantity selection unit, increases the weight of the language model for the acoustic analysis frame without the selection flag, performs speech recognition processing, and obtains the speech recognition result. Output.

  According to the acoustic analysis frame reliability calculation apparatus of the present invention, since the value of a function using an outlier score is obtained as a measure of reliability in an acoustic analysis frame unit, which is the minimum unit when handling audio data, erroneous recognition The section can be identified with high accuracy.

  Also, the acoustic model adaptation apparatus of the present invention adapts the acoustic model using the acoustic feature quantity series with frame reliability output from the acoustic analysis frame reliability calculation apparatus of the present invention. The acoustic model can be efficiently adapted.

  In addition, since the speech recognition apparatus according to the present invention performs speech recognition processing by increasing the weight of the language model for the acoustic analysis frame to which the selection flag is not assigned, the speech recognition accuracy can be improved.

The figure which shows the function structural example of the acoustic analysis frame reliability calculation apparatus 100 of this invention. The figure which shows the operation | movement flow of the acoustic analysis frame reliability calculation apparatus 100. The figure which shows the correspondence of an audio | voice signal and a frame reliability scale. The figure which shows the function structural example of the acoustic analysis frame reliability calculation apparatus 200 of this invention. The figure which shows the function structural example of the acoustic model adaptation apparatus 300 of this invention. The figure which shows the function structural example of the speech recognition apparatus 400 of this invention.

  Embodiments of the present invention will be described below with reference to the drawings. The same reference numerals are given to the same components in a plurality of drawings, and the description will not be repeated.

  FIG. 1 shows an example of the functional configuration of an acoustic analysis frame reliability calculation apparatus 100 according to the present invention. The operation flow is shown in FIG. The acoustic analysis frame reliability calculation device 100 includes an initial acoustic model 10, a speech recognition unit 11, and a frame confidence measure calculation unit 12. The acoustic analysis frame reliability calculation device 100 is realized by reading a predetermined program into a computer including, for example, a ROM, a RAM, a CPU, and the like, and executing the program by the CPU.

  The speech recognition unit 11 divides the speech signal into acoustic analysis frames having a predetermined time length, extracts acoustic feature amounts in units of the acoustic analysis frames, and performs speech recognition using the acoustic feature amounts and the initial acoustic model 10. Then, the maximum likelihood phoneme ID is assigned to the acoustic analysis frame and output as a phoneme ID / acoustic likelihood acoustic feature quantity sequence in which the acoustic feature quantity, acoustic likelihood, and phoneme ID are associated with each other (step S11). Here, the audio signal is, for example, a digital signal that has been made discrete at a sampling frequency of 10 kHz, and is assumed to have a length of several tens of minutes to several hours. The acoustic analysis frame has a time length of 10 ms obtained by collecting 100 discrete values, for example. Assuming that the time length of the acoustic analysis frame is 10 ms, for example, an acoustic feature amount of 180000 frames is extracted from an audio signal of 30 minutes.

  The acoustic feature quantity is a real-valued vector, and any method such as LPC cepstrum or MFCC may be used as an extraction method. The phoneme ID is an identifier that associates a phoneme of text information obtained by voice recognition of a voice signal and an acoustic analysis frame.

  FIG. 3 illustrates the relationship between the acoustic analysis frame and the phoneme ID. The first line is an audio signal, the second line is a speech recognition result, the third line is an acoustic feature series, the fourth line is a phoneme ID, the fifth line is an acoustic likelihood, and the sixth and seventh lines are frames described later. A confidence measure and a selection flag. This example, a speech signal as "Hello", the recognition result is "Konniki is (koNnikiwa)" shows a case.

  The phoneme ID is a phoneme of text information obtained as a speech recognition result, and in this example, each phoneme of (k, oN, n, i, k, i, w, a) is calculated using the initial acoustic model 10. It is a phoneme with the largest acoustic likelihood. The phoneme ID is associated with each acoustic analysis frame of the acoustic feature amount series. “K” for the first to third acoustic analysis frames, “o” for the fourth to ninth acoustic analysis frames, and “N”, “n”, “i” thereafter. , “K”, “i”, “w”, “a” are assigned phoneme IDs. As described above, the acoustic feature quantity sequence with phoneme ID / acoustic likelihood is composed of respective time series in which the acoustic feature quantity, the phoneme ID, and the acoustic likelihood are associated with each other.

  For example, reference 1 (Masayoshi et al. “Free Speech Recognition Technology“ VoiceRex ”that Listens to Natural Conversations with Customers” ”, NTT Technical Journal, Vol. 18, No. .11, pp.15-18, 2006.) can be obtained by the conventional speech recognition technology.

  The frame confidence measure calculation unit 12 receives the phoneme ID / acoustic likelihood-equipped acoustic feature quantity sequence output from the speech recognition unit 11 and classifies the acoustic feature quantity for each phoneme ID to create an acoustic feature quantity set. Outlier detection is performed on each acoustic feature amount set to obtain an outlier score, a function value using the outlier score is obtained as a measure of reliability, and the value of the function is calculated for each acoustic analysis frame. The assigned acoustic feature quantity series with frame reliability is output (step S12, FIG. 2).

  The process of outputting the acoustic feature quantity series with frame reliability is repeated until an audio signal is not input for a certain period of time or until an operation stop signal (not shown) is input to the control unit 15 (in step S15). No). The control unit 15 performs the control of the time series operation and the control of the operation end in steps S11 and S12. The function of the control unit 15 is not a special technical feature of this embodiment but a general one.

The frame confidence measure is a value indicating the probability of association between a certain acoustic analysis frame and a phoneme ID. There is an acoustic analysis frame that is associated with an incorrect phoneme ID due to erroneous recognition that occurs in the speech recognition unit 11. For example, k shown in italics for the 27th to 29th characters in FIG.
Is an acoustic analysis frame associated with. A frame confidence measure is used to detect such acoustic analysis frames.

  The frame confidence measure is calculated based on the idea that “acoustic feature quantities that are outliers cannot be trusted when acoustic feature quantities associated with the same phoneme are collected”. Specifically, the procedure is as follows.

  First, each acoustic analysis frame of the phoneme ID / acoustic likelihood acoustic feature quantity series is classified for each associated phoneme ID, and a set of acoustic feature quantities for each phoneme ID is created.

  Next, outlier detection is performed on a set D of acoustic feature values of a certain phoneme ID (X), and an outlier score is assigned to each acoustic feature value. As a method of calculating the outlier score of each vector from a set of real-valued vectors, existing methods such as LOF (Local Outlier Factor) and 1 class SVM (One Class Supprt Vector Machine) are used.

For example, the outlier score LOF (d _i ) when LOF is used is calculated by the following equation.

Here, d _i is the i-th acoustic feature and is included in the set D (1 ≦ i ≦ M). Further, k is an integer equal to or greater than 1, and is a parameter when calculating the outlier score LOF (d _i ). Usually, a value of about k = 10 to 20 is used and given in advance as a constant.

N _k (d _i ) of the denominator is a set in which the acoustic feature amount closest to d _i to the acoustic feature amount closest to the k th in the set D is collected. The Euclidean distance is used as the distance. | N _k (d _i ) | is the number of acoustic feature quantities included in N _k (d _i ), and is usually | N _k (d _i ) | = k.

  Lrd (x) representing the density of data around the acoustic feature quantity x is calculated by the following equation.

  Here, kdist (x) is a distance between a certain acoustic feature quantity x and the kth closest acoustic feature quantity, and dist (x, y) is a distance between the acoustic feature quantity x and y. An outlier when 1 class SVM is used is a distance from the class boundary surface.

Next, the weighted sum of the acoustic likelihood and the outlier score assigned to each acoustic feature amount is used as the frame confidence measure of each acoustic analysis frame. The frame confidence measure C _t is calculated by the following equation, where L _{t is} the acoustic likelihood and O _t is the outlier score. The frame confidence measure C _t is a function value defined as a measure of reliability. The frame confidence measure _Ct is given for each acoustic analysis frame as shown in the sixth line of FIG.

Here, α is an acoustic likelihood weight. α is a real value between 0 and 1, and when set to 0, the outlier score is the frame confidence measure as it is, and when it is set to 1, the acoustic likelihood L _t is the frame confidence measure as it is. Usually, it is set to about 0.5. The acoustic likelihood weight α may be set as a constant in the frame confidence measure calculation unit 12 in advance, or may be given from the outside as indicated by a broken line in FIG.

  As described above, according to the acoustic analysis frame reliability calculation device 100, the value of a function defined as a measure of reliability can be obtained in units of acoustic analysis frames, which is the minimum unit for handling audio data. That is, it becomes possible to detect the erroneous recognition section in units of acoustic analysis frames.

  In addition, about the speech recognition part 11 which outputs phoneme ID and the acoustic feature-value series with acoustic likelihood, another structure can also be considered. Next, the operation of the acoustic analysis frame reliability calculation device 200 having the voice recognition unit 11 as the voice recognition unit 21 will be described.

  FIG. 4 shows a functional configuration example of the acoustic analysis frame reliability calculation device 200 of the present invention. The acoustic analysis frame reliability calculation device 200 is obtained by replacing the speech recognition unit 11 of the acoustic analysis frame reliability calculation device 100 (FIG. 1) with a speech recognition unit 21.

  The voice recognition unit 21 includes a voice recognition unit 211 and an acoustic feature amount alignment unit 212. The voice recognition unit 211 recognizes the voice signal using the initial acoustic model and outputs a voice recognition text. The voice recognition unit 211 can be configured by the conventional voice recognition technique disclosed in the above-described Reference Document 1.

  The acoustic feature amount alignment unit 212 receives the voice signal and the voice recognition text output from the voice recognition unit 211 as input, divides the voice signal into acoustic analysis frames having a predetermined time length, and acoustic features of each acoustic analysis frame. An acoustic feature quantity sequence is generated by extracting the quantity. And the phoneme series corresponding to the said acoustic feature-value series is acquired from the input speech recognition text, and the initial acoustic model 10 is used among the correspondence between each frame of the acoustic feature-value series and each phoneme of the phoneme series. The correspondence with the largest acoustic likelihood calculated in this way is selected to associate each frame of the acoustic feature quantity sequence with the phoneme, and a phoneme ID is assigned to each acoustic analysis frame.

  That is, the acoustic feature amount alignment unit 212 determines how many frames each phoneme of the phoneme sequence determined by the speech recognition text continues. The acoustic likelihood of each acoustic analysis frame in correspondence with the largest acoustic likelihood is also given to each frame at the same time, and each time series is given to the frame confidence measure calculation unit 12 as an acoustic feature quantity sequence with phoneme ID and acoustic likelihood. Is output.

  As described above, even in the configuration including the speech recognition unit 21, the value of the function defined as a measure of reliability can be obtained in the unit of the acoustic analysis frame which is the minimum unit when handling speech data, and the erroneous recognition section is detected. The accuracy can be improved. The acoustic model adaptation apparatus 300 and the speech recognition apparatus 400 using the acoustic analysis frame reliability calculation apparatuses 100 and 200 described above are also conceivable. Next, these devices will be described.

[Acoustic model adapting device]
FIG. 5 shows a functional configuration example of the acoustic model adaptation apparatus 300 of the present invention. The acoustic model adaptation device 300 includes acoustic analysis frame reliability calculation devices 100 and 200, a feature amount selection unit 316, and an acoustic model adaptation unit 317.

  The acoustic analysis frame reliability calculation devices 100 and 200 are the acoustic analysis frame reliability calculation device 100 (FIG. 1) or the acoustic analysis frame reliability calculation device 200 (FIG. 4). The feature quantity selection unit 316 receives an acoustic analysis frame with frame reliability output from the acoustic analysis frame reliability calculation devices 100 and 200 as an input, and selects an acoustic analysis frame having a reliability for each acoustic analysis frame equal to or greater than the selection threshold θ. An acoustic feature quantity series with a selection flag to which a selection flag indicating selection is given is output.

  The selection flag is shown in the bottom line of FIG. “◯” in the figure indicates that the acoustic analysis frame is selected. “X” indicates that the acoustic analysis frame is not selected. In the example shown in FIG. 3, it can be seen that a selection flag of “x” is given to the 27th to 29th acoustic analysis frames in which “Chi” is misrecognized as “Ki”, and is not selected.

  The selection threshold θ may be set in advance in the feature quantity selection unit 316 as a constant, or may be given to the feature quantity selection unit 316 from the outside. If the selection threshold θ is set to a small value, more acoustic analysis frames are selected, but the number of erroneously recognized acoustic analysis frames is also increased. On the other hand, if the selection threshold θ is set to a large value, it is difficult to select the erroneously recognized acoustic analysis frames, but the total number of selected acoustic analysis frames is reduced.

  The selection threshold θ may be set to a fixed value based on experience, or may be obtained by automatic calculation so as not to select the lower 10% of the frame confidence measure, for example. In order to obtain the selection threshold θ so as not to select an arbitrary percentage, the average value μ and the standard deviation σ of all frame confidence measures are obtained, and the selection threshold θ is automatically calculated based on a statistical method. good.

  The acoustic model adaptation unit 317 receives the acoustic feature amount series with the selection flag output from the feature amount selection unit 316 as an input, and uses the acoustic feature amount of each acoustic analysis frame without the selection flag as a statistic to be calculated. The updated statistic is obtained by multiplying a real value non-selection weight of 0 or more and 1 or less, the parameters of the initial acoustic model 10 are updated based on the updated statistic, and output as an after-adaptation acoustic model. The initial acoustic model 10 is internal to the acoustic analysis frame reliability calculation apparatuses 100 and 200, but the acoustic model may be adapted using another base acoustic model.

  In updating the parameters of the acoustic model, formulas for calculating parameters after updating existing methods such as the maximum likelihood estimation method and the maximum posterior probability method are used. In the equation for updating the parameters of the acoustic model, a statistic calculated using the acoustic feature amount of each acoustic analysis frame is multiplied by a non-selection weight w in the case of an unselected acoustic feature amount.

  The non-selection weight w is a real value of 0 or more and 1 or less. If it is set to 0, an acoustic parameter not selected is not used at all, and the updated parameter is calculated. On the other hand, if it is set to 1, the same updated parameter as that of normal acoustic model adaptation ignoring the non-selection flag is calculated. Further, if it is set between 0 and 1, the updated parameter is calculated with the influence of the unselected acoustic feature amount reduced. The non-selection weight w is normally set to a value of about 0 to 0.5.

  Specifically, when using the maximum likelihood estimation method, reference 2 (Koichi Shinoda, “Speaker Adaptation Technology for Speech Recognition Using Probabilistic Models”. D-II, Information / System, II-Pattern Processing, J87- D-II (2), 371-386, 2004-02-01.) The updated calculation of the average and variance parameters of the acoustic model is changed to the following equation.

The changes in the equation are that the statistic (within sigma) regarding the acoustic feature value x _t of each acoustic analysis frame is multiplied by the weight w _t of the t-th acoustic analysis frame, and the denominator of the equations (5) and (6). The total number of frames T is replaced with T _select + wT _reject . The value of the weight w _t of the t-th acoustic analysis frame is “1” when the t-th acoustic analysis frame is selected, and the input non-selection weight w when it is not selected. T _select is the total number of selected acoustic analysis frames, T _reject is the total number of unselected acoustic analysis frames, and T _select + T _reject = T. T _select + wT _reject is the total number of frames in which the number of unselected acoustic analysis frames is estimated. In the formula (6), 'means transposition.

  When the maximum posterior probability method is used, the calculation formula of the average and variance parameters of the updated acoustic model described in Reference 2 is changed to the following formula.

As in the case of the maximum likelihood estimation method, the change in this case is that the statistic regarding the acoustic feature amount x _t of each acoustic analysis frame is multiplied by the weight w _t of the t-th acoustic analysis frame, and the total frame of the denominator. The number T is replaced by T _select + wT _reject .

  The acoustic model adaptation apparatus 300 assigns a selection flag indicating whether or not the feature selection unit 316 uses the acoustic feature series for adaptation of the acoustic model, and the acoustic model adaptation unit 317 correctly excludes the erroneous recognition section. Acoustic model adaptation is performed using feature quantities. Therefore, a high adaptation effect is obtained in the acoustic model adaptation unit 317, and the post-adaptation acoustic model can achieve high speech recognition accuracy.

[Voice recognition device]
FIG. 6 shows a functional configuration example of the speech recognition apparatus 400 of the present invention. The speech recognition apparatus 400 includes acoustic analysis frame reliability calculation devices 100 and 200, a feature amount selection unit 316, a speech recognition unit 418, an acoustic model 419, and a language model 420. As is clear from the reference numerals, the acoustic analysis frame reliability calculation devices 100 and 200 and the feature amount selection unit 316 are the same as the acoustic model adaptation device 300 described above.

The speech recognition unit 418 outputs a speech recognition result using the acoustic model 419 and the language model 420 as an acoustic feature amount series input with a selection flag output from the feature amount selection unit 316, and is given a selection flag. For the acoustic analysis frames that are not, the speech recognition process is performed by increasing the weight of the language model 420. In other words, since the acoustic feature amount of the acoustic analysis frame to which the selection flag is not assigned is likely to contain an error, the calculation of the score that depends on the acoustic feature amount is handled lightly. In short, the weight of the language model 420 is set larger than the weight of the acoustic model 419, the score of the acoustic analysis frame is calculated, and the speech recognition process is performed. As a result, improvement in voice recognition accuracy can be expected.

  According to the acoustic analysis frame reliability calculation devices 100 and 200 described above, the value of a function using an outlier score is obtained as a measure of reliability in units of acoustic analysis frames, which is the minimum unit for handling audio data. Thus, it is possible to accurately identify the erroneous recognition section. Since the acoustic model adaptation device 300 including the acoustic analysis frame reliability calculation devices 100 and 200 performs the acoustic model adaptation using the acoustic feature quantity sequence with the frame reliability, the cost of collecting data of the adaptation data is reduced, The acoustic model can be adapted efficiently. In addition, since the speech recognition apparatus 400 performs speech recognition processing by increasing the weight of the language model for an acoustic analysis frame to which no selection flag is assigned, speech recognition accuracy can be improved.

  When the processing means in the above apparatus is realized by a computer, the processing contents of the functions that each apparatus should have are described by a program. Then, by executing this program on the computer, the processing means in each apparatus is realized on the computer.

  The program is distributed by selling, transferring, or lending a portable recording medium such as a DVD or CD-ROM in which the program is recorded. Further, the program may be distributed by storing the program in a recording device of a server computer and transferring the program from the server computer to another computer via a network.

  Each means may be configured by executing a predetermined program on a computer, or at least a part of these processing contents may be realized by hardware.

Claims (6)

The audio signal is divided into acoustic analysis frames having a predetermined time length, and acoustic features are extracted in units of the acoustic analysis frames, and acoustic likelihood is calculated using the acoustic features and the initial acoustic model, and the acoustic analysis is performed. A speech recognition unit that outputs a phoneme ID / acoustic likelihood acoustic feature quantity sequence in which a maximum likelihood phoneme ID is assigned to a frame and the acoustic feature quantity, acoustic likelihood, and phoneme ID are associated with each other;
Using the phoneme ID / acoustic likelihood acoustic feature series as an input, classifying the acoustic feature quantity for each phoneme ID to create an acoustic feature quantity set, and detecting outliers for each acoustic feature quantity set To obtain an outlier score for each acoustic feature amount , obtain a function value using the outlier score as a measure of reliability, and assign a value of the function to each acoustic analysis frame. A frame confidence measure calculation unit for outputting a sequence of attached acoustic features,
An acoustic analysis frame reliability calculation device comprising: In the acoustic analysis frame reliability calculation device according to claim 1,
The voice recognition unit
Speech recognition means for recognizing a speech signal using an initial acoustic model and outputting speech recognition text;
The speech signal and speech recognition text are input, the speech signal is divided into acoustic analysis frames having a predetermined time length, and acoustic feature amounts of each acoustic analysis frame are extracted to generate an acoustic feature amount sequence. A phoneme sequence is acquired from the recognized text, and a phoneme having the maximum acoustic likelihood is assigned to each acoustic analysis frame as a phoneme ID using the initial acoustic model, and the acoustic feature value, the acoustic likelihood, and the phoneme ID are associated with each other. An acoustic feature quantity alignment means for outputting a series of acoustic feature quantities with phoneme ID / acoustic likelihood;
An acoustic analysis frame reliability calculation device characterized by comprising: The acoustic analysis frame reliability calculation device according to claim 1 or 2,
A selection flag indicating that an acoustic analysis frame having a reliability of each acoustic analysis frame equal to or higher than a selection threshold is selected by using the acoustic feature amount series with frame reliability output from the acoustic analysis frame reliability calculation device as an input. A feature quantity selection unit that outputs the selected acoustic feature quantity series with a selection flag;
Using the acoustic feature quantity sequence with the selection flag as an input, the statistic calculated using the acoustic feature quantity of each acoustic analysis frame not assigned with the selection flag is multiplied by a non-selection weight of a real value between 0 and 1. An updated acoustic model, and updates the parameters of the initial acoustic model based on the updated statistics and outputs an adapted acoustic model; and
An acoustic model adaptation apparatus comprising: The acoustic analysis frame reliability calculation device according to claim 1 or 2,
A selection flag indicating that an acoustic analysis frame having a reliability of each acoustic analysis frame equal to or higher than a selection threshold is selected by using the acoustic feature amount series with frame reliability output from the acoustic analysis frame reliability calculation device as an input. A feature quantity selection unit that outputs the selected acoustic feature quantity series with a selection flag;
A speech recognition unit that receives the acoustic feature quantity sequence with the selection flag as input, increases the weight of the language model for the acoustic analysis frame without the selection flag, and performs speech recognition processing and outputs a speech recognition result;
A speech recognition apparatus comprising: The speech recognition unit divides the speech signal into acoustic analysis frames of a predetermined time length, extracts acoustic feature quantities in units of the acoustic analysis frames, and calculates acoustic likelihood using the acoustic feature quantities and the initial acoustic model. A speech recognition process for outputting a phoneme ID / acoustic likelihood-equipped acoustic feature quantity sequence in which a maximum likelihood phoneme ID is assigned to the acoustic analysis frame and the acoustic feature quantity, acoustic likelihood, and phoneme ID are associated with each other; ,
The frame confidence measure calculation unit generates the acoustic feature amount set by classifying the acoustic feature amount for each phoneme ID, using the phoneme ID / acoustic likelihood acoustic feature amount sequence as input, and generating each acoustic feature amount Outlier detection is performed on the set, an outlier score is obtained for each acoustic feature amount, a function value using the outlier score is obtained as a measure of reliability, and the function is calculated for each acoustic analysis frame. A frame confidence measure calculation process for outputting an acoustic feature quantity sequence with a frame reliability to which a value is assigned;
Acoustic analysis frame reliability calculation method including Functions of each part of the acoustic analysis frame reliability calculation device according to claim 1, the acoustic model adaptation device according to claim 3, and the speech recognition device according to claim 4 are provided to a computer. A program to be executed.

Images